Blog
How AI agents complete multi-step tasks
AI agents complete multi-step tasks with a loop: a planner turns the goal into concrete steps, an executor carries out one step at a time using tools such as web search and a browser, and each step ends with a reported result. Good agents add checkpoints for human approval, budgets so a stuck step stops, and routing so simple steps run on cheaper models.
Published
By Pebs by NoxThe loop, step by step
- 01
Plan
The goal becomes a short list of steps. In Pebs you read this plan before anything happens.
- 02
Execute
The executor works one step with tools: live web, the Cloud Browser, files, memory.
- 03
Finish the step
A step can only end with an explicit finish and a deliverable, not a description of having done it.
- 04
Report and continue
The result is shown to you; the next step starts. Follow-ups reopen the same task.
Where thinking levels come in
Not every step deserves the strongest model. In GHI-1, routine steps such as signing in, entering a verification code or watching a page run on Vela 2 and move up to Orion Flash 2 if they get stuck. Harder steps run on the model for your chosen level, from Orion Flash 2 at Low to Kimi K3 at Max.
Checkpoints and budgets
- Approval checkpoints before sends, purchases, submissions and external changes.
- Secure cards for anything secret, so the model never sees it.
- An action budget per step. A step that loops is asked to finish with what it has, then stopped.
- Resumption: tasks continue after sleeps, approvals and your own takeover of the browser.
Why deliverables matter
A multi-step agent that only reports progress leaves you to redo the work. Pebs ends each step with the list, the summary with sources, the file or the confirmation itself, so the output of a task is something you can use.
Pebs by Nox
Give it something to do.
Coming to Nox on all platforms November 1, 2026. Sign up for Nox to join the waitlist.